Re: [RFC PATCH 1/5] mm: intorduce __GFP_UNMAPPED and unmapped_alloc()

[Kent Overstreet]

On Thu, May 18, 2023 at 12:03:03PM -0700, Song Liu wrote:
> On Thu, May 18, 2023 at 11:47 AM Song Liu <song@xxxxxxxxxx> wrote:
> >
> > On Thu, May 18, 2023 at 10:24 AM Kent Overstreet
> > <kent.overstreet@xxxxxxxxx> wrote:
> > >
> > > On Thu, May 18, 2023 at 10:00:39AM -0700, Song Liu wrote:
> > > > On Thu, May 18, 2023 at 9:48 AM Kent Overstreet
> > > > <kent.overstreet@xxxxxxxxx> wrote:
> > > > >
> > > > > On Thu, May 18, 2023 at 09:33:20AM -0700, Song Liu wrote:
> > > > > > I am working on patches based on the discussion in [1]. I am planning to
> > > > > > send v1 for review in a week or so.
> > > > >
> > > > > Hey Song, I was reviewing that thread too,
> > > > >
> > > > > Are you taking a different approach based on Thomas's feedback? I think
> > > > > he had some fair points in that thread.
> > > >
> > > > Yes, the API is based on Thomas's suggestion, like 90% from the discussions.
> > > >
> > > > >
> > > > > My own feeling is that the buddy allocator is our tool for allocating
> > > > > larger variable sized physically contiguous allocations, so I'd like to
> > > > > see something based on that - I think we could do a hybrid buddy/slab
> > > > > allocator approach, like we have for regular memory allocations.
> > > >
> > > > I am planning to implement the allocator based on this (reuse
> > > > vmap_area logic):
> > >
> > > Ah, you're still doing vmap_area approach.
> > >
> > > Mike's approach looks like it'll be _much_ lighter weight and higher
> > > performance, to me. vmalloc is known to be slow compared to the buddy
> > > allocator, and with Mike's approach we're only modifying mappings once
> > > per 2 MB chunk.
> > >
> > > I don't see anything in your code for sub-page sized allocations too, so
> > > perhaps I should keep going with my slab allocator.
> >
> > The vmap_area approach handles sub-page allocations. In 5/5 of set [2],
> > we showed that multiple BPF programs share the same page with some
> > kernel text (_etext).
> >
> > > Could you share your thoughts on your approach vs. Mike's? I'm newer to
> > > this area of the code than you two so maybe there's an angle I've missed
> > > :)
> >
> > AFAICT, tree based solution (vmap_area) is more efficient than bitmap
> > based solution.

Tree based requires quite a bit of overhead for the rbtree pointers, and
additional vmap_area structs.

With a buddy allocator based approach, there's no additional state that
needs to be allocated, since it all fits in struct page.

> > First, for 2MiB page with 64B chunk size, we need a bitmap of
> >      2MiB / 64B = 32k bit = 4k bytes
> > While the tree based solution can adapt to the number of allocations within
> > This 2MiB page. Also, searching a free range within 4kB of bitmap may
> > actually be slower than searching in the tree.
> >
> > Second, bitmap based solution cannot handle > 2MiB allocation cleanly,
> > while tree based solution can. For example, if a big driver uses 3MiB, the
> > tree based allocator can allocate 4MiB for it, and use the rest 1MiB for
> > smaller allocations.

We're not talking about a bitmap based solution for >= PAGE_SIZE
allocations, the alternative is a buddy allocator - so no searching,
just per power-of-two freelists.

> 
>  Missed one:
> 
> Third, bitmap based solution requires a "size" parameter in free(). It is an
> overhead for the user. Tree based solution doesn't have this issue.

No, we can recover the size of the allocation via compound_order() -
hasn't historically been done for alloc_pages() allocations to avoid
setting up the state in each page for compound head/tail, but it perhaps
should be (and is with folios, which we've generally been switching to).